CRNN-Transformer：基于混合神经网络的音乐风格分类方法

CRN-Transformer: A Music Style Classification Method Based on Hybrid Neural Networks

摘要: 在数字化时代背景下，音乐信息检索技术的发展日新月异，音乐风格分类作为该领域的核心任务之一，对于提升音乐推荐系统的性能和用户体验具有重要意义。为了从复杂的音频信号中准确识别和分类音乐风格，本研究设计并开发了一种基于混合神经网络的CRNN-Transformer模型。本文的技术创新集中在基于CNN算法引入的残差神经网络模块(RESNET)、双向门控循环神经单元(GRU)和Transformer模块的关键改进上。首先，采用ResNet模块来增强模型在频谱空间特征提取的能力，通过残差连接解决深层网络中的梯度消失问题；其次，引入双向GRU模块以更好地捕捉时序信息，通过同时考虑过去和未来的信息，进一步提升模型对序列数据的理解；最后，集成Transformer模块，利用自注意力机制建模长距离依赖关系，从而增强模型的表示能力。本研究使用音频的梅尔频率倒谱系数(MFCC)作为输入特征，进行特征提取和时序建模。实验结果表明，相比于传统的CNN网络，CRNN-Transformer分别在F1-score，Precision，Recall三个指标上提升了14.8%，16%，13.7%，而在与其他主流模型进行的比较中，各指标也均取得了最佳表现。

Abstract: In the context of the digital era, the development of music information retrieval technology is changing with each passing day. As one of the core tasks in this field, music style classification is of great significance to improve the performance and user experience of music recommendation system. In order to accurately identify and classify music styles from complex audio signals, this study designed and developed a CRNN-Transformer model based on hybrid neural networks. The technical innovation of this paper focuses on the key improvements of residual neural network module (RESNET), bidirectional gated recurrent neural unit (GRU) and Transformer module based on CNN algorithm. Firstly, the ResNet module is used to enhance the ability of the model to extract features in the spectrum space, and the gradient disappearance problem in the deep network is solved by residual connection. Secondly, the bidirectional GRU module is introduced to better capture the timing information, and the model’s understanding of sequence data is further improved by considering both past and future information. Finally, the Transformer module is integrated, and the self-attention mechanism is used to model long-distance dependencies, thereby enhancing the representation ability of the model. In this study, Mel-Frequency Cepstral Coefficients (MFCC) of audio are used as input features for feature extraction and time series modeling. The experimental results show that compared with the traditional CNN network, CRNN-Transformer improves F1-score, Precision and Recall by 14.8%, 16% and 13.7%, respectively. In comparison with other mainstream models, each index also achieves the best performance.